
A Verb-based Algorithm for Multiple-Relation Extraction from Single Sentences Qi Hao1, Jeroen Keppens1, and Odinaldo Rodrigues1 1Department of Informatics, King’s College London, London, United Kingdom Abstract— With the growing amount of unstructured ar- performance of proposed algorithm were also compared with ticles written in natural-language, automated extracting single-relation extraction algorithm. knowledge of associations between entities is becoming The ultimate goal of relation extraction is to construct essential for many applications. In this paper, we develop networks from text data that indicate various associations automated verb-based algorithm for multiple-relation ex- among different entities across different areas. For example, traction from unstructured data obtained on-line. Named the sentence "The quality of magnesium status directly Entity Recognition (NER) techniques were applied to extract influences the Biological Clock function (BC)" contains a biomedical entities and relations were recognized by algo- relation between magnesium and Biological Clock function. rithms with Natural Language Processing (NLP) techniques. The first step is to enable computers to process words and Evaluation based on F-measure with random sample of sentences through Natural Language Processing (NLP) tech- sentences from biomedical literature results an average niques. Several open libraries and toolkits were developed precision of 90% and recall of 82%. We also compared the recently, such as Stanford’s CoreNLP [1] and OpenNLP performance of proposed algorithm with single-relation ex- [2]. They provide a rich set of common NLP tools such traction algorithm, indicating improvements of this work. In as tokenization, lemmatization, part-of-speech (POS) tag- conclusion, the preliminary study indicates that this method ging, parsing, etc. In addition, two essential steps – entity for multiple-relation extraction from unstructured literature recognition and relationship extraction – have recently had is effective. With different training dataset, the algorithm can tremendous progress. Existing Named Entity Recognition be applied to different domains. The automated method can (NER) tools can recognize not only general terms such be applied to detect and predict hidden relationships among as proper nouns, but also more specific entities such as varying areas. diseases and symptoms [3], [4], [5], [6]. As for relation extraction, five main methods are currently used: extraction Keywords: Multiple-relation extraction, Natural Language Pro- based on co-occurrence, link-based extraction, extraction cessing (NLP), Named Entity Recognition, verb-based algorithm using machine learning approaches, rule-based extraction and verb-based extraction [7], [8]. The first four of these 1. Introduction methods can deal with simple relations between two entities with some target words with relatively low precision and A substantial amount of valuable knowledge is recorded recall. Verb-based methods on the other hand normally have in the form of unstructured text data, such as news, emails, higher precision and can be applied in a variety of domains. journal articles and conference papers. The biomedical liter- However, current verb-based approaches can only extract a ature is one body of knowledge of this form. But although single relation embedded in a sentence composed of a verb text documents provide an effective way to disseminate phrase sandwiched between two entities of interest. If the knowledge within a relatively small community or narrow sentence contains multiple relations, then existing verb-based field of study, it becomes very hard or impossible for algorithms can only extract one of these relations. More humans to fully comprehend all the knowledge comprised details about these five methods are discussed in Section 5. in this form within much larger communities or within In this work, we propose a verb-based algorithm using collections of related disciplines and specialties. This pa- existing NLP techniques and NER tools to extract relations per contributes to ongoing efforts to develop mechanisms from text, including multiple relations embedded in the for automated knowledge extraction from texts written in same sentence. Data is automatically downloaded and then natural language. In this work, we propose a verb-based processed using standard NLP techniques to extract the algorithm to extract multiple relationships between entities entities. Subsequently, instead of identifying target verb from unstructured articles written in natural language. The phrases by POS tagging and parsing alone, we extract verb algorithm is evaluated by an experiment based on biomedical relations using semantically similar verbs. Single-relation literature to extract bio-entities and relations between them extraction algorithms work well when extracting simple co- including substance, symptom, disease and body part. The occurrence relations such as Entity-Verb-Entity. However, we enhanced this process so that it can deal with three Data Pre-processing common sentence structures which embed multiple relations Sentence within a single sentence. By analyzing the structure of the Sentence Pos Tagging with POS Parsing from text tagging clauses, our algorithm is able to extract multiple verb-based relations connected by a relative pronoun such as which or Sentence with POS that. Similarly, an analysis of the sentence level conjunctive tagging and structure, allows the algorithm to extract multiple verb-based Structure relations connected by conjunctions such as and or but. Finally, by analysing the phrase level conjunctive structure, NER entities Relation Relations the algorithm is also able to extract one-to-many or many- extraction to-one relations. UMLS, Detect In our experiments, our algorithm achieved an average WordNet, verbs VerbNet verbs precision of 90% and a recall rate of 82%. It is worth mentioning that our algorithm is not restricted to a fixed Text processing for information extraction domain or a particular set of verbs. Fig. 1: Overview of a single iteration of the extraction This rest of the paper is organised as follows. Section 2 process presents a brief review of the relevant NLP techniques and NER tools used to pre-process the texts. Section 3 explains the proposed algorithm and experiments. Section 4 provides on gene entities recognition using the GENETAG corpus an evaluation of the proposed algorithm. In Section 5 we for training. MetaMap is a highly developed software to discuss some related work and this is followed by some map biomedical text to the UMLS Metathesaurus or equiv- conclusions and future work in section 6. alently, to recognize and extract Metathesaurus concepts in biomedical publications. Abner has multiple models for 2. Background recognizing protein, DNA, RNA, cell line and cell type In order to process and analyse texts written in natural with the NLPBA and BioCreative corpora. Based on those language, NLP techniques should be introduced. NLP is a existing approaches, the proposed methodology is discussed cross disciplinary field in artificial intelligence, and compu- in Section 3. tational linguistics iinvestigating ways to enable computers to interact with humans and understand human natural lan- guages. Some standard techniques include word segmenta- 3. Methodology tion, POS tagging, word sense disambiguation, parsing, and NER. Word segmentation enables computers identifying and The main objective of the proposed research is to extract extracting valid words from a continuous stream. POS tag- relationships between entities from biomedical publications. ging helps computers classify words into categorize such as The input of the system is a set of publication texts re- noun, verb, and adjectives to each word. Parsing determines garding some particular biomedical entities. The texts are structure of the sentences based on POS tags. NER helps analyzed using standard NLP techniques and entities and computers locate and classify named entities that are rigid verb relations are recognized and extracted by the proposed designators [9] in natural texts into pre-defined categories algorithm. such as proper names of persons, organizations, and certain The system starts by identifying and extracting PubMed biological species and substances. publication abstract records that contain the target biomed- Many open source toolkits and libraries have been de- ical terms such as "magnesium deficiency", "migraine at- veloped these days. OpenNLP is a machine learning based tack". Those records are stored in text files with their corre- toolkit that has been widely used for standard NLP tasks sponding PubMed ID (PMID), title and abstract. Separated [10], [11], [12]. OpenNLP provides a command line script sentences from texts are the input of one iteration of the and an API as well. It can also be used as a package in Java algorithm. An overview of a single iteration process is shown program or R program. in Figure 1. The proposed algorithm is divided into two Although OpenNLP can perform simple NER tasks such main tasks: Data pre-processing and text processing. For data as recognizing person and company names, locations or pre-processing, NLP techniques are applied in the selected times, NER remains a crucial and complex task for biomed- sentences for POS tagging and parsing. For text processing, ical domain due to the complexity of bio-entities and lack NER are applied for extracting entities and verb-detection of human-annotated data. Many
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages7 Page
-
File Size-